The first provisions of the EU AI Act are already in effect, and it is no accident that the rules around forbidden AI practices took priority. You can read more about the broader framework in our EU AI Act Summary for full context on how these bans fit into the regulation as a whole.
Article 5 of the EU AI Act, which defines what AI uses are strictly prohibited, entered into force on February 2, 2025, significantly before the rest of the AI Act provisions. According to the European Commission's 2025 AI Act Implementation Report, over 60% of EU Member States have already designated national competent authorities responsible for enforcing the prohibited practices provisions (European Commission, 2025).
In April 2025, the European Commission also published non-binding Guidelines on Prohibited AI Practices, offering detailed legal interpretations, practical examples and illustrative case studies.
As that date is already behind us, organizations and developers should be fully aware of which AI behaviors are no longer allowed, whether due to how they function or what they are used for. For those managing compliance programs, strong AI governance is essential to identify and prevent these high-risk uses.
What Are the EU AI Act Prohibited Practices?
TL;DRThe EU AI Act bans eight categories of AI use that pose unacceptable risks to fundamental rights. These prohibitions took effect on February 2, 2025, making them the first enforceable obligations under the Act. Violations carry fines of up to €35 million or 7% of global annual turnover. All organizations deploying or developing AI within the EU must verify their systems do not fall into any banned category.
The EU AI Act identifies eight specific types of AI use that are entirely forbidden. These represent the highest tier of risk under the Act's classification system: "unacceptable risk." Unlike high-risk AI systems, which are permitted under strict conditions, prohibited practices cannot be deployed under any circumstances (with narrow exceptions for law enforcement biometric identification).
Manipulating Vulnerabilities to Alter Behavior
TL;DRAI systems that target a person's specific vulnerability, such as age, disability, or economic hardship, to alter their behavior in a harmful way are banned. The prohibition covers any system designed to exploit a known weakness to push someone toward a decision they would not otherwise make, regardless of whether the harm is financial, psychological, or physical.
Subjects affected by the AI Act must not develop, offer, or use AI systems that exploit a person's specific vulnerability to influence their behavior in a way that could harm them or someone else.
Such vulnerabilities include:
- Belonging to a particular age group
- Having a disability
- Being in a distinct social or economic situation
A real-world example of this prohibited practice is any marketing strategy that intentionally singles out vulnerable individuals to direct them toward specific actions. Consider an AI-driven advertising tool that monitors online activity and detects users whose web searches often include "affordable mobility aids" and "home care tips", indicating that such users might be older adults with mobility challenges. If the AI starts targeting these users with ads for overpriced, unverified health supplements or home devices that offer little real benefit, such a practice violates the EU AI Act: it exploits an individual's physical or psychological vulnerabilities to drive a harmful decision.
Deceptive or Manipulative Techniques
TL;DRAI systems using subliminal, deceptive, or manipulative techniques to distort a person's decision-making are prohibited. The ban targets methods that operate below conscious awareness and lead individuals toward choices they would not freely make. This includes artificially manufactured urgency, hidden persuasion patterns, and covert behavioral nudging deployed through AI interfaces.
The AI Act includes a prohibition aimed at preventing negative impacts on an individual through deceptive, manipulative, or other subliminal techniques implemented in an AI system.
The reason why these methods are prohibited lies in the fact that they undermine an individual's ability to make a fully informed choice. They manipulate a person's decision-making and provoke behaviors that rarely reflect free will.
For example, imagine a travel booking website using an AI system to monitor a user's browsing behavior. Based on the information collected, the system detects when someone is urgently searching for last-minute flights. At that moment, the system displays a pop-up with a message: "Hurry, only 2 seats left at this price!" Even if there are far more seats available, this tactic pressures the traveler into booking a ticket that may not be needed or affordable. The AI has manufactured urgency to override rational decision-making.
Social Scoring Leading to Negative Outcomes
TL;DRAI-driven social scoring that classifies individuals based on their behavior or personal traits and then penalizes them with unfavorable treatment is banned. The prohibition covers systems that produce scores leading to outcomes disproportionate to the original context, or that treat people unjustly based on aggregated personal data. Both public and private sector uses fall within the ban.
Using AI to classify individuals based on their actions or traits in a way that may assign them a negative "social score" is not permitted.
AI-driven scoring can easily undermine fundamental rights and principles such as fairness, equality, and dignity. That is why the Act prohibits it outright.
Consider an AI system used by a rental platform to analyze users' online reviews and social media activity. If the AI assigns a "low trust" score to a potential lessee based on past complaints or minor negative comments, and as a result refuses their application or demands a significantly higher deposit, this would be discriminatory. It penalizes individuals for behavior unrelated to their tenancy, violating their basic rights.
Predicting Criminal Behavior Through Profiling
TL;DRAI systems that predict an individual's likelihood of committing a crime based solely on profiling, personality traits, or social interactions are prohibited. The ban protects against algorithmic bias and presumption of guilt without factual evidence. AI tools that assist investigators using objective, documented evidence remain permitted under the Act.
AI systems must not be used to foresee the likelihood of an individual committing a crime based solely on their social interactions or personality traits.
Such a prohibited system could, for example, use the results of an individual's personality test, or scan someone's social media profile, and on that basis conclude that a particular person may potentially commit a crime. Those conclusions rest on subjective criteria rather than proven facts. They are prone to errors, bias, and severe privacy violations.
This ban does not apply to AI tools that only assist human investigators using objective, provable information. When an AI system supports a law enforcement officer by highlighting documented, evidence-based leads, rather than speculating on criminality risk based on personality, it is permitted under the EU AI Act. The use of AI in enforcement can be properly directed and limited, rather than prohibited entirely.
Inferring Emotions of Employees or Students
TL;DREmotion-detection AI systems deployed in workplaces or educational settings are banned under the Act. In these contexts, individuals are typically in subordinate positions, making emotion inference prone to discrimination and unfair treatment. The prohibition does not extend to medical or safety applications where emotion detection serves a protective, clinically supervised purpose.
Any AI system aimed at assessing the emotions of employees or students is forbidden under the AI Act. Emotion-detection tools in work or educational settings can introduce bias and lead to unfair treatment.
Consider an AI application that records team members during online meetings to evaluate their emotional states. Such a tool falls under this prohibition. In a workplace or educational setting, individuals are typically in a subordinate position relative to the other party, which makes emotion-detection systems prone to fostering discrimination and unjust treatment.
Not all emotion-detection AI is prohibited. The AI Act permits systems used for medical or safety purposes, where such tools can be beneficial, provided they operate outside high-risk scenarios such as employment or schooling.
Untargeted Data Scraping for Facial Recognition
TL;DRBuilding facial recognition databases through untargeted scraping of images from the internet or CCTV footage is prohibited. The ban applies regardless of whether the images were publicly available or captured for legitimate surveillance purposes. This provision aligns with GDPR principles on lawful, proportionate, and transparent data processing.
The EU AI Act prohibits the use of AI to collect considerable amounts of images or recordings of individuals for building facial recognition databases.
Data scraping has been a legally controversial topic for some time, and the AI Act deepens that debate. The ban applies to all sources of visual data, regardless of whether the data is widely available online or captured via CCTV cameras. It applies even to recordings made for legitimate security or monitoring purposes, such as private property surveillance: the collected data cannot be repurposed to feed a facial recognition database without the data subject's consent.
The justification behind this restriction is to restrain unauthorized processing of personal data and protect individuals' privacy rights. By aligning with GDPR principles, the AI Act confirms that any collection of biometric data must remain lawful, proportionate, and transparent, regardless of the original collection context.
Discriminatory Biometric Categorization
TL;DRAI systems that sort or label individuals based on biometric data to infer characteristics such as race, political opinions, religious beliefs, trade union membership, sex life, or sexual orientation are prohibited. The ban prevents discriminatory profiling through facial analysis or other biometric techniques. Legitimate law-enforcement biometric applications under judicial authorization remain permitted.
AI systems may not sort or label people based on characteristics such as race, political opinions, trade union membership, religious beliefs, sex life, or sexual orientation.
For example, if a retail store installs cameras equipped with AI to analyze customers' facial features and estimate their race, and the system then uses that inferred information to show certain groups higher prices or different product recommendations, the system violates the AI Act's prohibition.
On the other hand, legitimate biometric applications, including those used by law enforcement for identifying suspects under judicial authorization, remain permitted. Using biometric data for legally supervised purposes, such as locating a known offender, is allowed under the EU AI Act.
Real-Time Remote Biometric Identification in Law Enforcement
TL;DRReal-time facial recognition in public spaces by law enforcement is banned except in three narrow circumstances: finding trafficking victims, preventing imminent threats to life, and locating suspects of serious crimes carrying at least four years' imprisonment. Even where exceptions apply, a fundamental-rights impact assessment, national authorization, and registration in the EU high-risk AI database are all required.
Real-time remote biometric identification systems are banned because they pose serious risks to individual rights and privacy.
This form of biometric surveillance is one of the most debated aspects of the EU AI Act. Deploying facial recognition at large gatherings, such as concerts or mass protests, to identify attendees and cross-check them against criminal databases is not allowed under ordinary circumstances.
A narrow list of circumstances may justify such systems:
- Detecting human trafficking or abduction victims
- Preventing a highly credible threat to public safety
- Locating a person suspected of a serious crime punishable by at least four years imprisonment
Even when one of these exceptions applies, all of the following additional conditions must be met before the system can be used:
- Authorities must assess the severity, likelihood, and scope of harm that could arise if the system is not deployed.
- A fundamental-rights impact assessment (FRIA) must be conducted to understand the potential consequences of deploying the system.
- The system must be registered in the EU database of high-risk AI systems.
- National-level approval is required: the responsible authority in the EU member state must confirm that use of the system is both necessary and proportionate, based on objective evidence.
- Both the national market-surveillance authority and data-protection authority must be notified. Those authorities must send annual reports to the European Commission detailing the system's use.
If an EU Member State chooses to permit real-time biometric identification in public spaces more broadly, it may do so through national legislation, provided that the national law fully complies with the AI Act's restrictions.
As the EU AI Act prohibited practices have taken effect, organizations must remain alert in identifying and eliminating banned AI uses. By clearly defining and enforcing these prohibitions, the EU AI Act seeks to safeguard fundamental rights and promote responsible innovation.
Over time, we will learn whether these rules strike the right balance between protection and progress. In the meantime, businesses should keep auditing their AI systems and stay informed on the evolving guidance in order to maintain compliance and build a trustworthy AI ecosystem. A structured AI governance approach is key to ongoing risk management and regulatory alignment. For organizations that need to classify their AI systems, maintaining an up-to-date AI inventory is a practical first step. Violations of these prohibitions carry the highest tier of EU AI Act penalties, with fines reaching €35 million or 7% of global annual turnover.
For providers of high-risk AI systems, the next obligation after classification is the conformity assessment procedure.
Further Reading on Whisperly
Questions & Answers
What are the prohibited AI practices under the EU AI Act?+
The EU AI Act bans eight categories of AI use: exploiting vulnerabilities, deceptive or manipulative techniques, social scoring, criminal behavior prediction through profiling, workplace or school emotion detection, untargeted facial recognition data scraping, discriminatory biometric categorization, and real-time remote biometric identification by law enforcement. These prohibitions took effect on February 2, 2025. For the full regulatory context, see the EU AI Act Summary.
When did the EU AI Act prohibited practices take effect?+
The prohibited practices provisions under Article 5 of the EU AI Act entered into force on February 2, 2025. This was the first set of obligations to become enforceable, ahead of the broader Act's phased rollout. Organizations should already have completed an AI inventory to verify none of their systems fall into a banned category.
What are the penalties for violating the EU AI Act prohibited practices?+
Violations of the prohibited practices provisions carry the highest fines under the Act: up to €35 million or 7% of total worldwide annual turnover from the preceding financial year, whichever is higher. SMEs pay the lower of the two amounts. Full details on the tiered penalty structure are available in our guide to EU AI Act penalties.
Are there any exceptions to the real-time biometric identification ban?+
Yes. Three narrow exceptions exist: searching for trafficking or abduction victims, preventing an imminent threat to life or a foreseeable terrorist attack, and locating suspects of serious crimes punishable by at least four years' imprisonment. Even when an exception applies, a fundamental-rights impact assessment and national authorization are mandatory. Systems must also be registered in the EU database of high-risk AI systems.
Does the EU AI Act prohibit all emotion-detection AI?+
No. The Act prohibits emotion-detection AI in workplaces and educational institutions because of the inherent power imbalance and risk of discrimination. However, emotion-detection systems used for medical or safety purposes remain permitted. Organizations should conduct a conformity assessment to determine whether their system falls within the ban.
How does the EU AI Act define manipulative AI techniques?+
The Act defines manipulative AI as any system using subliminal, deceptive, or manipulative techniques that distort a person's behavior and lead them toward decisions they would not otherwise make. This includes artificially manufactured urgency, hidden persuasion patterns, and covert behavioral nudging. A strong AI governance framework helps organizations detect and prevent these practices.
What is the difference between prohibited practices and high-risk AI under the EU AI Act?+
Prohibited practices are AI uses banned outright because they pose unacceptable risks to fundamental rights, such as social scoring or manipulating vulnerabilities. High-risk AI systems are permitted but subject to strict compliance requirements including conformity assessments, transparency obligations, and human oversight. Learn more about classification in our guide to high-risk AI systems.
How can organizations check if their AI systems are banned under the EU AI Act?+
Start by creating a comprehensive AI inventory of all AI systems in use. Map each system against the eight prohibited categories in Article 5 of the Act. For systems that do not fall into a banned category but may be high-risk, proceed with a conformity assessment. An AI governance platform can help systematize this process.
Reviewed by: Tamara Zavisic, AI Governance Specialist